config.yaml

log_dir: ./Models/Finetune
save_freq: 1
log_interval: 10
device: cuda
epochs: 50 
batch_size: 2
max_len: 160 # maximum number of frames
pretrained_model: ./Models/Finetune/base_model_120k_vi.pth
load_only_params: true # set to true if do not want to load epoch numbers and optimizer parameters
debug: false

data_params:
  train_data: data_22k/train_phn.txt
  val_data: data_22k/dev_phn.txt
  root_path: data_22k
  n_speakers: 152
symbol: #Total 189 symbols
  pad: "$"
  punctuation: ';:,.!?¡¿—…"«»“” '
  letters: "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
  letters_ipa: "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
  extend: "∫̆ăη͡123456"

preprocess_params:
  sr: 24000
  spect_params:
    n_fft: 2048
    win_length: 1200
    hop_length: 300

training_strats:
  #All modules: 'decoder', 'predictor', 'text_encoder', 'style_encoder', 'text_aligner', 'pitch_extractor', 'mpd', 'msd'
  freeze_modules: [''] # Not updated when training.
  ignore_modules: ['spk_emb', 'spk_ln', 'style_ln', 'gate'] # Not loading => fresh start. IMPORTANT: 'text_aligner' and 'pitch_extractor' are util pretraineds DO NOT ignore them.

model_params:
  dim_in: 64 
  hidden_dim: 512
  max_conv_dim: 512
  n_layer: 3
  n_mels: 80
  max_dur: 50 # maximum duration of a single phoneme
  style_dim: 128 # style vector size
  dropout: 0.2

  ASR_params:
    input_dim: 80
    hidden_dim: 256
    n_layers: 6
    token_embedding_dim: 512

  JDC_params:
    num_class: 1
    seq_len: 192

  # config for decoder
  decoder: # either hifigan or istftnet or vocos

      type: hifigan
      resblock_kernel_sizes: [3,7,11]
      upsample_rates :  [10,5,3,2]
      upsample_initial_channel: 512
      resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
      upsample_kernel_sizes: [20,10,6,4]
      
      # type: 'istftnet'
      # resblock_kernel_sizes: [3,7,11]
      # upsample_rates :  [10, 6]
      # upsample_initial_channel: 512
      # resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
      # upsample_kernel_sizes: [20, 12]
      # gen_istft_n_fft: 20
      # gen_istft_hop_size: 5

      # type: 'vocos' 
      # intermediate_dim: 1536
      # num_layers: 8
      # gen_istft_n_fft: 1200
      # gen_istft_hop_size: 300
  
loss_params:
    lambda_mel: 5. # mel reconstruction loss
    lambda_gen: 1. # generator loss
    
    lambda_mono: 1. # monotonic alignment loss (TMA)
    lambda_s2s: 1. # sequence-to-sequence loss (TMA)

    lambda_F0: 1. # F0 reconstruction loss
    lambda_norm: 1. # norm reconstruction loss
    lambda_dur: 1. # duration loss
    lambda_ce: 20. # duration predictor probability output CE loss

optimizer_params:
  lr: 0.0001 # general learning rate
  ft_lr: 0.00001 # learning rate for acoustic modules